Scored across 16 tools
Every tool carries an explicit 'Not this tool for' section that routes the adjacent cases (generate_molecule vs batch_generate vs compare_molecules, lookup vs lookup_molecule_data, predict_spectrum vs generate_spectrum, titration_curve vs species_distribution vs calculate_ph). The closest pairs (lookup/lookup_molecule_data, calculate_ph/generate_titration_curve) are separated by a clear 'text fact vs drawn figure' rule. The only non-agent-facing tool, save_png, is explicitly marked 'never call this directly'.
The set is overwhelmingly consistent snake_case verb_noun: generate_molecule, generate_reaction, generate_spectrum, compare_molecules, predict_spectrum, calculate_ph, export_anki_deck. Minor deviations are batch_generate (noun_verb order, inverted relative to the other generate_* tools) and the bare verb lookup, which lacks a noun object while its sibling is lookup_molecule_data.
Sixteen tools is slightly above the comfortable band, but the domain genuinely spans structure drawing, reaction schemes, mechanisms, spectra, TLC, 3D, acid-base calculations and data lookup, so each tool maps to a real task. The count is inflated by one internal helper (save_png) that agents should never call, which is the only clearly non-earning entry.
The surface covers the full drawing lifecycle: single, batch, comparison, 3D, reactions, mechanisms, spectra (draw and predict), TLC, three acid-base calculators, data lookup and Anki export — a wide, coherent workflow. Gaps are acknowledged rather than hidden: mechanisms exist only for SN1/SN2/Fischer esterification, and there is no SMILES-to-name conversion or reaction prediction, but agents can route those to generate_reaction or lookup.
